The speed and computational accuracy of modern digital computers are well-known. However, all digital computers solve problems in a sequential fashion through the use of numerical computation. While the processing unit contained in a simple pocket calculator can easily out-perform the human brain in number crunching tasks, digital computers are able to accomplish this sophisticated numerical analysis only on a step-by-step basis. Digital computers exhibit their best abilities when presented with a serially programmable algorithm. Digital computers are not capable of sophisticated parallel processing, such as that required when a human undertakes the task of pattern recognition. Problems such as comparing the fingerprint found at the scene of a crime with a data base full of fingerprints is the sort of practical and necessary problem that arises and yet is not easily solved by a digital computer. To the extent that digital computers have been programmed to match the fingerprint found at the scene of the crime with an existing fingerprint in the files, lengthy serial searches of memory are required to digitally achieve accurate pattern recognition.
A matrix algebra based on an associative memory model was described by J. J. Hopfield in his paper "Neural Networks and Physical Systems with Emergent Collective Computational Abilities," proceedings of the National Academy of Science U.S.A., 1982, Vol. 79, pp. 2554-2558. The Hopfield model utilizes feedback and nonlinear thresholding to force the output pattern to be the stored pattern which most closely matches an input pattern presented to the associative memory system. A digital emulation of this model requires large storage and computational effort for the manipulation of an association matrix used in the model. For example, in order to store two-dimensional image patterns consisting of N.times.N pixels, the model requires a matrix with N.sup.4 entries be used.
A natural implementation of an associative memory model would be one which uses optical technology. Optical associative memory systems store information as patterns so that, upon the introduction of a stored pattern to the system, the system is able to recall the stored pattern and perform a match. These optical systems achieve massive parallel processing. The ability of an optical associative memory to perform such a function has wide application in the fields of pattern recognition and image understanding. Used in conjunction with a laser beam, specially treated photosensitive film or plates act as holograms. A hologram is a frozen "picture" of an object wherein the image of the object is recorded on the film plate as an interference pattern between a reference beam of plane waves (which is directed only at the photographic film) and an object wave front (which is created by reflection from the object, where the object wave front is made by the same coherent source that produced the reference beam). Holograms are characterized as having extremely good spatial coherence. The light used to produce the hologram, normally laser beam, exhibits a high degree of temporal coherence. In order to view the recorded holographic image, one redirects coherent light along the same path as the reference beam which originally recorded the hologram. A viewer views the hologram along the same line of sight that connected the object and the hologram during its recording. Directing a new reference beam on the hologram causes an image to appear which, in a lensless environment, gives rise to a three-dimensional image. The lifelike dimensionality of a lensless image produced in a hologram is due to the fact that, unlike a photograph, a hologram stores not only amplitude changes but also records phase changes as interference fringes resulting from the interaction between spatially coherent object and reference beams.
Holograms are characterized by very precise and lifelike imaging. In addition, a hologram, when viewed from different angles, produces different views of the recorded image. The hologram is programmable for use in storing a plurality of images, by varying the angle of the reference beam used to record the image. The information stored within a hologram is recorded throughout the holographic medium; even a portion of the hologram retains the complete record. It therefore can be seen that holograms are quite useful in parallel processing systems. Furthermore, holograms are inherently useful for optical pattern recognition mechanisms.
Among the types of holograms known in the art are the volume, Fresnel, and Fraunhofer holograms. The volume holograms have a thickness and can be used to record either amplitude or phase modulated images without the generation of both primary and conjugate waves that is inherent with thin holograms. Fraunhofer holograms are characterized as holograms that record distant objects. Larger and closer positioned objects produce Fresnel holograms.
The Fourier transform hologram uses a lens and is adaptable for memory storage purposes. As is well known in the numerical analysis arts, the Fourier transform is a mathematical tool wherein any function may be broken up into a sum of sinusoidal superimposed patterns. This manner of dividing a function into its Fourier components is known as defining the Fourier transform of a function. In Fourier transform holography, an object wavefront holographically captured, after it has undergone a Fourier transformation. To do this, a photographic holographic plate is placed at the back focal plane of the lens. A flat object, such as a transparency, is placed at the same distance in front of the lens as the photographic plate is behind it. The object's wave front, when it reaches the plate, has been Fourier-transformed by the lens. The holographic pattern produced as an image is quite unlike the original object. If the object is illuminated only by coherent light, such as a laser beam, and if a reference beam is provided at an angle to the plate, the Fourier transform can be recorded as a hologram.
Pattern recognition has used Fourier transform holograms in another fashion to perform the operation of convolution. The best way to understand convolution is to look at an example. If one were to convolve a first transparency having three dots with a second transparency having one triangle, using a holographic Fourier transform, one obtains three triangles, one at each position of the dots. A related operation mathematically similar to convolution is correlation. The result of correlating two identical objects is a sharp peak at a position corresponding to the shift value which superimposes the two objects. The peak is greatly reduced if the two objects are not identical, making correlation useful in pattern recognition.
To correlate two transparencies (also referred to as objects) one simply positions a first object one focal length in front of a lens and a Fourier transform hologram of a second object one focal length back of this lens. A second lens is positioned in back of the Fourier transform hologram of the second object. The correlation of the first and second objects appears one focal length behind the second lens.
An example of optical pattern recognition using correlation would be where in a printed page of text, one could recognize a particular word or letter at some position on a page. Wherever the particular word appears in the text, a bright spot of light highlights the word in the correlated image. Wherever the word occurs on the page, there will be a corresponding bright spot of light in the correlated image called a correlation peak. Thus, the nature of holograms and lenses combined in an optical system using a coherent light source allows the operation of pattern recognition to occur. Such a device has been characterized as an optical neural computer. The term "neural" is derived from the fact that the parallel processing of a hologram to provide an associative memory is similar to that of a human brain's neural system in that the stored information is not localized.
Heretofore, one such optical associative memory has been proposed by Abu-Mostafa and Psaltis in Scientific American, vol. 256, no. 3 in an article entitled "Optical Neural Computers," at page 88 (March, 1987). In that article an optical thresholding device and a pinhole array were used as part of a double hologram associative memory system.
The applicants have previously disclosed (as co-inventors) an associative memory system in U.S. Pat. No. 4,739,496, entitled "Associative Holographic Memory Apparatus Employing Phase Conjugate Mirrors", which issued on Apr. 19, 1988 on patent application Ser. No. 06/786,884, filed Oct. 11, 1985. Also, the applicants are co-inventors in U.S. Pat. No. 4,750,153, "ASSOCIATIVE HOLOGRAPHIC MEMORY APPARATUS EMPLOYING PHASE CONJUGATE MIRRORS IN A TWO-IMAGE WAVE MIXING CONTRA-DIRECTIONAL COHERENT IMAGE AMPLIFIER", issued on June 7, 1988 on patent application Ser. No. 06/821,237, filed Jan. 22, 1986. (The disclosures contained in both applications are hereby incorporated by reference.) Hughes Aircraft company, the assignee of this application, is also the assignee of these two pending applications. These systems also employ primarily all-optical elements.
As indicated above, optical elements, such as the hologram, make excellent associative memory storage devices. When a distorted input image is presented to a system which includes at least one hologram (containing a clear representation of that image), the system processes light through its components in such a manner as to correlate and match the distorted input image with one of the images stored on the hologram. The sharper the correlation peaks, the better the match. All optical systems are excellent parallel processors but generally may not be shift-invariant and furthermore, they may exhibit optical and gain losses in the system as the image is processed. In order to achieve a good match, an optical associative memory must have good thresholding and gain so that the correlation peak which reconstructs the reference beam (when the image is to be reconstructed) is sharp and bright. Losses of light intensity in the system are inevitable as the light is processed through an optical system as disclosed in the above-incorporated applications or as that disclosed in the Abu-Mostafa article, supra. Additionally, reconstruction and phase conjugation of the reference beam in the all-optical systems described in U.S. Pat. No. 4,739,496 and U.S. Pat. No. 750,153 is achieved inherently by use of phase conjugate mirrors, (PCMs) using for example BaTiO.sub.3 material. In such systems, thresholding is determined by physical processes in the PCMs and is neither easily alterable nor readily adjustable. Also, such optical systems heretofore have required at least a second for the PCM to respond. BaTiO.sub.3 --based optical techniques are relatively slow, in a computer sense. Phase conjugate mirrors of an all-optical component system may be used to fully reconstruct and return an image to its point of origin to achieve pattern recognition. Nonlinearities in the phase conjugate mirrors are used to select those stored objects which exceed a threshold, based on the overlap of computed integrals of the object input with the stored objects. Although, experimentally, store-and-recall of two objects with shift invariance, was achieved, the gains achieved by phase conjugate mirrors were not enough to overcome hologram losses. Additionally, the non-linearities of the phase conjugate mirrors were difficult to control.
It is therefore an object of this invention to provide a system which makes use of the pattern recognition properties of a hologram but in such a manner that optical losses are kept to a minimum, thresholding with gain achieved, and sharp correlation of images at the hologram accomplished, with shift invariance. U.S. Pat. Nos. 4,546,248 and 4,556,986, both issued to Glenn D. Craig and assigned to the United States (NASA), disclose electro-optical systems used to process images with incoherent light sources. The systems represent attempts to vary spatially the optical gain of signals without thresholding or enhancement of optical images. Such references show the state of the electro-optical art, but do not in themselves advance the achievement of the objects of this invention to provide an associative memory system.
Liquid crystal light valves (LCLVs) suitable for use in the present invention include, for example, those shown in the following U.S. patents all assigned to Hughes Aircraft Company, the assignee of the present invention:
U.S. Pat. No. 3,824,002, "Alternating Current Liquid Crystal Light Valve", issued to T. D. Beard, on July 16, 1974; U.S. Pat. No. 4,019,807, "Reflective Liquid Crystal Light Valve with Hybrid Field Effect Mode", issued to Boswell et al. on Apr. 26, 1977;
U.S. Pat. No. 4,018,509, for "Optical Data Processing system with Reflective Liquid Crystal Light Valve", issued to Boswell et al. on Apr. 19, 1977;
U.S. Pat. No. 4,378,955, for "Method and Apparatus for a Multimode Image Display with a Liquid Crystal Light Valve", issued to Bleha et al. on Apr. 5, 1983;
U.S. Pat. No. 4,239,348, "High Resolution AC Silicon MOS--Light Valve Substrate" issued to J. Grinberg et al. on Dec. 16, 1980;
U.S. Pat. No. 4,443,064 "High Resolution AC Silicon MOS--Light Valve Substrate" issued to J. Grinberg et al. on Apr. 17, 1984; and,
U.S. Pat. No. 4,127,322, "High Brightnesss Full Color Image Light Valve Projection System", issued to Jacobson et al., on Nov. 28, 1978.